MARS workup Β· 2026-07-04

Three companies,
one database.

A pure-private venture, a public-market alt-asset behemoth, and one that just crossed over. Same shape, same joins, same substrate β€” shown against the tables MARS has been building.

MARS's job is visibility into companies the public market hasn't seen yet. The harder questions are continuity β€” does the substrate carry a company cleanly across the private-to-public crossover? β€” and scale β€” does it hold up on a firm that does dozens of billion-dollar deals a year? The three cases below answer both. Every number came from a single SELECT against mars.companies_v2 joined through the v2 canonical layer. No hand entry.

01 The private-only case

Databricks

Twelve years of SEC Reg D substrate on a $60B+ private company β€” visible only because we ingest Form D directly.

Rounds
19
Series A 2014 through Series L 2025
Raised
$111 B
Equity + debt facilities, per canonical extraction
M&A
10
MosaicML, Tabular, Neon
Form D filings
18
2013 β†’ 2025 continuous
Officers
142
Linked person_id
Real estate
1
Office acquisition Β· Oct 2025
Databricks has never filed a public 10-K. It has filed eighteen Form Ds. MARS reads them, extracts issuer, offering size, and every named officer, resolves each officer to a canonical person_id, and threads the whole arc together β€” from the $15M 2013 filing to the $10B 2025 mega-offering that matched the Series L round.

Form D β€” the private timeline in filings

Directly from mars.form_d_filings_v2

Filed Offering size Sold Round it funded
2013-09-25$15 M$14 MSeries A
2014-07-03$33.7 M$33.4 MSeries B
2017-08-22$140 M$140 MSeries D
2019-02-07$250 M$250 MSeries E
2019-11-05$400 M$400 MSeries F
2021-02-02$1.0 B$1.0 BSeries G
2021-09-15$1.6 B$1.6 BSeries H
2023-09-29$1.22 B$1.18 BSeries I (dual filing)
2023-11-13$1.0 B$685 MSeries I follow-on
2025-01-02$10.0 B$8.59 BSeries J mega round
2025-09-23$1.74 B$1.62 BSeries K (dual)
2025-12-31$5.0 B$4.08 BSeries L

Twelve of eighteen shown. The whole file is WHERE company_id = '32c37…' β€” same UUID resolves every Form D filing against the same canonical company record.

Recent mega rounds and their investor lists

JSONB investor arrays canonicalized to investors_v2

Date Round Size Investors
2026-02-09 Equity + Debt $5.0 B Goldman Sachs, Morgan Stanley, JPMorgan, Neuberger Berman, Qatar Investment Authority, Glade Brook Capital
2025-12-16 Series L $4.0 B Andreessen Horowitz, Insight Partners, Fidelity, JPMorgan AM, BlackRock, Blackstone, Coatue, GIC, MGX, NEA, Ontario Teachers, Robinhood Ventures, T. Rowe Price, Temasek, Thrive Capital, Winslow Capital
2025-09-09 Series K $1.0 B Andreessen Horowitz, Insight Partners, MGX, Thrive Capital, WCM Investment Management

M&A β€” the acquisitions Databricks made

Same company_id on the acquirer side of merger_deals_v2

Announced Target Value Note
2026-06-16Panther Labsundisc.Cloud security analytics
2025-05-14Neon$1.0 BServerless Postgres β€” vs Snowflake Iceberg push
2025-02-04BladeBridgeundisc.Legacy-warehouse migration
2024-06-04Tabular$1.75 BIceberg format co-creator; bought hours after Snowflake talks broke down
2023-10-23Arcion$100 MReal-time replication
2023-07-19MosaicML$1.3 BGenerative-AI training platform; launched Databricks's foundation-model story
2023-05-03Okeraundisc.Governance / access control

Officers β€” the SEC-visible governance

Named on Form D filings across the twelve years above and resolved through form_d_officers_v2 to canonical persons_v2 rows.

Name Roles Note
Ali GhodsiExecutive Officer Β· Director Β· PromoterCEO
Ion StoicaExecutive Officer Β· Director Β· PromoterCo-founder Β· Berkeley AMPLab
Matei ZahariaExecutive Officer Β· Director Β· PromoterCo-founder Β· created Spark
Ben HorowitzDirectorAndreessen Horowitz
David ConteExecutive OfficerCFO
Elena DonioDirectorBoard
Jonathan ChadwickDirectorBoard Β· former VMware CFO

Sample of 142 officer rows resolved to canonical persons. Extractor typos β€” Ali Ghosdi for Ali Ghodsi β€” get queued through the Qwen β†’ Haiku β†’ Grok-web resolution tier chain and cleared without human intervention. Same person_id on either side of the merge.

02 The scale case

Blackstone

The world's largest real-estate investor doing dozens of billion-dollar deals across property, credit, and M&A β€” all threaded together by the reit vertical MARS scaffolded two days ago.

Real-estate deals
54
$262.8 B tracked
M&A deals
63
Cross-industry
Funding events
20
As an issuer (funds & BDCs)
Property types
7
Data ctr, industrial, resi, hotel…
Deal types
7
Acq, sale, financing, JV, offering…
Source
reit
gdelt_metadata Β· re-gdelt-v1
$262.8 billion of real-estate activity across 54 individual transactions, all inferred from public news through the reit vertical's gdelt_metadata pipeline. Blackstone doesn't file a Form D for these β€” none needed at the parent β€” so the substrate is entirely built up from headlines the extractor read and classified.

Real-estate deals β€” by shape and size

From real_estate_deals_v2 Β· reit vertical, live since 2026-07-03

By deal type

TypeCountTotal
Property acquisition19$96.5 B
Financing9$74.9 B
REIT offering4$42.8 B
Fund raise3$22.1 B
Property sale16$15.5 B
Development JV2$10.0 B
Portfolio sale1$1.0 B

Largest 6 transactions

DateType Β· assetValue
2026-06-03Industrial acquisition$81.0 B
2026-06-01Residential financing$60.0 B
2026-05-05Data-center REIT offering$16.0 B
2026-04-14Data-center REIT offering$16.0 B
2026-03-27Fund raise$15.0 B
2026-01-22Data-center acquisition$10.0 B

The concentration in data centers is real β€” Blackstone owns QTS Realty, and the AirTrunk $16.1B acquisition (below) added Asia-Pacific footprint. Data-center deals cluster in real_estate_deals_v2 as property_type='data_center'; the M&A shows up separately in merger_deals_v2. Same company_id stitches both.

Largest M&A transactions

Where the RE strategy crosses into operating-company M&A

Announced Target Value Sector
2024-09-04AirTrunk$16.1 BData centers Β· Asia-Pacific
2022-10-31Copeland (formerly Emerson Climate)$14.0 BHVAC / industrial
2024-04-08Apartment Income REIT (AIR Communities)$10.0 BResidential
2024-04-29L'Occitane International S.A.$6.4 BConsumer
2025-02-24Safe Harbor Marinas$5.65 BMarina operator
2023-03-14Cvent Holding Corp.$4.6 BEvent tech

63 M&A rows total. The three duplicate rows on Apartment Income REIT β€” Apartment Income REIT Corp. / Apartment Income REIT (AIR Communities) / Apartment Income REIT Corp. (AIR Communities) β€” are exactly what the entity-resolution tier chain is built to collapse. Next Phase G fire, Qwen resolves them; if it can't, Grok-web does.

03 The private-to-public arc

CoreWeave

Same substrate carries the whole trajectory: 2019 seed Form D to 2025 IPO to $3.28B of insider Form 144 filings, all under one company_id.

Rounds
47
$3M seed 2019 Β· $20B NVIDIA 2026
M&A
7
Incl. Core Scientific $9B
Real estate
2
Data-center prop & financing
Form D filings
6
2019 β†’ 2024 (pre-IPO)
Officers
13
Linked person_id
Form 144 (post-IPO)
335
$3.28 B insider sales
From a $3 million Form D in March 2019 to $3.28 billion of Form 144 insider filings by June 2026. The officers named on the earliest SEC filings are the same names on the largest post-IPO exits β€” resolved by person_id, not by string matching.

Funding β€” the mega-round era

Selected billion-plus rounds from funding_deals_v2

Date Round Amount Notable investors
2026-05-19Debt + Equity$20.0 BNVIDIA Strategic
2026-05-18Delayed-Draw Term Loan$3.1 BMorgan Stanley, MUFG
2026-04-09Convertible Senior Notes$3.25 BPublic offering
2026-03-31Debt Facility$8.5 BGoldman Sachs, JPMorgan, Morgan Stanley, MUFG
2025-03-20IPO$2.41 BNASDAQ:CRWV
2024-03-31Debt Facility (pre-IPO)$8.0 BMUFG, Morgan Stanley, Goldman, JPMorgan

Acquisitions

Announced Target Value Note
2025-10-30Marimo Inc.undisc.AI-notebook workspace
2025-10-06Monolith AIundisc.UK physics-ML startup
2025-09-03OpenPipe Inc.undisc.Fine-tuning platform
2025-07-07Core Scientific$9.0 BPublic-to-public; Bitcoin miner β†’ AI DC footprint
2025-05-05Weights & Biasesundisc.MLOps platform
2023-01-25Conductor Technologiesundisc.Render farm; pre-IPO tuck-in

Post-IPO insider filings β€” same names, different table

Top sellers from form_144_v2 Β· resolved to persons_v2

Seller Role at issuer Filings Value
West Clay Capital LLCBrannin McBee vehicle32$938.5 M
Jack CogenDirector7$715.5 M
Brannin McBeeCo-founder, Executive Officer27$395.2 M
Michael IntratorCo-founder, CEO, Director28$266.4 M
Brannin J. McBee 2022 IRR TrustMcBee family trust7$258.2 M
Total across all 335 filings 335 $3.28 B

Michael Intrator and Brannin McBee both appear in form_d_officers_v2 from 2019 onward. The person_id on those officer rows is the same UUID resolved on the Form 144 sale rows β€” cross-linked at ingest, not stitched together at query time.

04 The through-line

What makes all three readable

The value isn't the count of rows. It's that everything joins cleanly.

One key, fourteen tables

Both companies live under a single companies_v2.company_id. Fourteen v2 tables carry FKs to it: funding, mergers, real estate, four SEC substrates, two officer bridges, deal articles, canonical persons, and the reit vertical's real_estate_deals_v2.

The join graph is discovered at runtime from Postgres foreign keys β€” new tables get picked up without code changes.

Direct-to-SEC ingest

Form D, 13D, 13G, 144, and ADV all land in v2 tables owned by mars. Officer parsing runs at ingest time, not on read. That's how Databricks has 142 canonical officer rows without a public 10-K, and how CoreWeave's pre-IPO officers are the same rows as its post-IPO insider sellers.

Deterministic quality gate

Every deal carries an integrity_tier: verified / vouched / warn. Consumer queries filter to the first two; the small "warn" tail is quarantined and auto re-audited. The $10B Databricks Form D and the $9B CoreWeave / Core Scientific merger both cleared the gate on ingest.

Tier chain closes the entity graph

Extractor typos like Ali Ghosdi, Biran Venturo, Michael Intrrator are auto-resolved through Qwen β†’ Haiku β†’ Grok-web. 891 canonical merges executed 2026-07-04 with 0 unique-constraint conflicts. The person named on a 2013 Form D stays the same person named on a 2026 Form 144.

How this was built. One SELECT against mars.companies_v2 for each subject. Every table above followed with a single JOIN. No manual entry, no external lookup β€” every row came from the same substrate that answers MCP queries, Metabase dashboards, and Raul's mergers pipeline.

What's underneath. Postgres mars schema, v2 canonical layer. Direct SEC ingest (adrian + kee, 2026-06-30) covers Form D / 144 / 13D / 13G / ADV. Phase G canonicalizes deals hourly. Entity resolution via the Qwen β†’ Haiku β†’ Grok-web tier chain. Real estate arrived via the reit vertical scaffolded 2026-07-02 and shipping 5,679 rows by 2026-07-04.

Reproducibility. Substitute either subject's company_id in the workup queries; get the same shape for any covered issuer. That's the promise.

MARS Β· 2026-07-04 Β· kee@kscope.io